Re: Which is simpler? (Was Re: [Suspend2-devel] Re: [ 00/10] [Suspend2] Modules support.)

[Nigel Cunningham]

Hi Andy.

On Tuesday 21 February 2006 14:32, Andy Lutomirski wrote:
> Matthias Hensler wrote:
> > Hi.
> >
> > On Mon, Feb 20, 2006 at 01:53:33AM +0100, Pavel Machek wrote:
> >>Only feature I can't do is "save whole pagecache"... and 14000 lines
> >>of code for _that_ is a bit too much. I could probably patch my kernel
> >>to dump pagecache to userspace, but I do not think it is worth the
> >>effort.
> >
> > I do not think that Suspend 2 needs 14000 lines for that, the core is
> > much smaller. But besides, _not_ saving the pagecache is a really _bad_
> > idea. I expect to have my system back after resume, in the same state I
> > had left it prior to suspend. I really do not like it how it is done by
> > Windows, it is just ugly to have a slowly responding system after
> > resume, because all caches and buffers are gone.
> >
> > I can only speak for myself, but I want to work with my system from the
> > moment my desktop is back.
>
> I Am Not A VM Hacker, but:
>
> What's the point of saving pagecache during suspend?  This seems like a
> total waste.  Why don't we save a list of pages in pagecache to disk,
> then, after resume, prefetch them all back in.  This will slow down
> resume (extra seeks, minimized if we sort the list, and inability
> to compress these pages), but it will speed up suspend, and it sounds
> a lot simpler.  There's already a patch to add swap prefetching, and
> this can't be much more complicated.

The page cache contains the process pages, among other things, so it can't all 
be discarded with impunity. You're right in suggesting that discarding them 
and then prefetching them would be a potential alternative, but it would 
actually be more complicated: you'd still have to remember which pages you 
wanted to fault back in, and some how store that info in the image. You'd 
also have to add code to do the faulting.  Right now, we don't (for the most 
part) care what a page is used for - we just save and reload it. I believe 
both implementations give you the option to set a soft upper limit on the 
size of the image, so you can still throw away cache if you want to.

> While I'm at it, here's another pie-in-the-sky idea.  If we had the
> ability to unmount an in-use filesystem (that lack is my single biggest
> pet peeve about Linux right now -- Windows has been able to do this for
> ages), then the process could be:
> 1. Atomic snapshot of userspace.  Snapshot all struct file's as well.
> 2. Unmount local filesystems.  Network filesystems probably can't be
> trashed by buggy suspend anyway.
> 3. Snapshot kernelspace.
> 4. Suspend happily without worrying about hosing filesystems, since
> they're all unmounted.
> 5. For extra safety, unmount anything that the suspend process mounted.
>
> -- Shutdown and restart --
>
> 6. Mount stuff and load the image, then unmount it (again using the
> in-use FS unmounting).
> 7. Restore the image and resume kernelspace.
> 8. Remount local filesystems and reattach all the struct files.
> 9. Resume userspace.  Start prefetching everything.
>
> This would seem to allow an ordinary program to handle image writing
> with no particular worries about disk access, except that unlinking
> files might confuse userspace (but not the FS) on resume.
>
> Feel free to tell me why this is impossible.  If no one tells me it's
> impossible, I may work on unmounting in-use filesystems and revoking
> fd's in a month when I have more free time.

It's not quite unmounting, but Suspend2 already freezes bdevs as part of the 
process of quiescing the system. This was the only way I could find to make 
XFS really complete pending I/O (and stop further submissions).

Regards,

Nigel

-- 
See our web page for Howtos, FAQs, the Wiki and mailing list info.
http://www.suspend2.net                IRC: #suspend2 on Freenode

Attachment: pgp00000.pgp
Description: PGP signature